Social Recommendations: Mentor and Leader Detection to Alleviate the Cold-Start Problem in Collaborative Filtering

Introduction

With the democratization of the Internet, users often need to be assisted in their search of information or search of items. Recommender systems have been proposed in the beginning of the 90's (Goldberg et al, 1992), with the aim to fulfill this need. Indeed, the volume of items that users can access is now so huge that they cannot get the information they want within a small amount of time; users are thus unsatisfied. This consequence can be dramatic for e-commerce services for example, that aim at increasing their sales and at developing customers' loyalty. As a consequence, recommender systems are increasing in popularity and are no more of secondary importance; they are becoming mandatory in many e-services.

Recommender systems are not simple information delivery systems; they recommend and display personalized information or pertinent items to users. They are a way to cope with the classical “one size fits all'' characteristic of many information delivery systems, such as classical search engines (Allan et al., 2003).

Recommender systems take into consideration the users' specific characteristics, represented under the form of users' profile (Adomavicius & Tuzhilin, 2005). An item is the minimal unit that a recommender system can manage. For example, an item can be a book, a movie, a web page, etc. Recommender systems are now exploited in many application domains, such as e-commerce (Paolino et al., 2009), e-learning (Zhuhadar et al., 2009), restaurants (Hosseini-Pozveh et al., 2009), news (Tintarev & Masthoff, 2006), etc.

Recommender systems generally fall into three categories: content-based systems which compute recommendations from the semantic content of items (Pazzani & Billsus, 2007); knowledge-based systems where recommendations rely on the knowledge about the domain, the users and pre-established heuristics (Burke et al, 1996); and at last collaborative filtering systems (Adomavicius & Tuzhilin, 2005) which compute recommendations by examining users' preferences on items.

The users' preferences managed by a collaborative filtering (CF) system are often expressed under the form of ratings and stored in users’ profiles. The structure of such a system can be represented under the form of a graph, with nodes being the users and links being the similarity of preferences among them. This graph can be viewed as a social network (Brun & Boyer, 2010), where the links are not social relations but preference relations. To compute recommendations for an active user a, a classical CF system exploits the known preferences of the users linked to a in the social network, as well as the values of the links.

In CF, a implicitly requests the preferences from his like-minded users about some items: he asks them for some recommendations. The ratings of the items a has not rated yet are then inferred from these recommendations. The items with the highest ratings are then recommended to a.

A collaborative filtering recommender system is thus a social process: not only the active user is involved in the recommendation process; other users are also. In CF, a's like-minded users are called his neighbors. Two main approaches are used to select a's neighbors: the memory-based approach and the model-based approach. In the memory-based approach, the set of neighbors is specific to each user; in the model-based approach, the set of neighbors can be specific to each class of users.

The search of the best set of neighbors has attracted much attention in the literature (Breese et al, 1998, Herlocker et al, 2004, Kim & Yang, 2007, Castagnos & Boyer, 2007). Classically the number of neighbors required to get high quality recommendations is about several dozens (Shardanand & Maes, 1995, Brun et al, 2009).

In the literature, the set of neighbors is only selected according to their similarity with the active user. These neighbors are then used as recommenders. However, we argue that these neighbors may be bad recommenders despite their similarity with the active user, since we do not consider their predictive capacity. As an example, they may be unable to recommend items that have not been rated by the active user a, if they have not rated them either. Thus, we claim that the use of the similarity as the only criterion for selecting neighbors is not sufficient.

In this book chapter, we propose a new approach of collaborative filtering: the delegate-based collaborative filtering. This approach has the following characteristics: